MCP Solved the Wrong Problem
Model Context Protocol is now universal AI infrastructure. But MCP solves connectivity, not accuracy — and those are not the same problem.
On March 11, 2026, three things happened that confirmed Model Context Protocol had graduated from emerging standard to universal infrastructure. Manufact raised $6.3M from Peak XV and YC to build open-source MCP tooling. Claude's Microsoft Office add-ins gained shared context and org-wide "Skills." And the server count crossed 10,000 public implementations, with ChatGPT, Gemini, Copilot, VS Code, AWS, GCP, Azure, and Cloudflare all in the ecosystem.
VentureBeat called it "the USB-C for AI." That framing is accurate. It also contains the problem that nobody in Wednesday's coverage mentioned.
What MCP actually solved
Before MCP, connecting an AI agent to your tools meant custom connectors. Slack integration. Salesforce integration. SharePoint integration. Each one built from scratch, each one its own maintenance headache. MCP replaced all of that with a single, consistent interface. AI agents can now talk to any software, database, or document system through one protocol.
That's a real infrastructure win. Enterprise AI deployment was genuinely slower and more expensive because of the connector fragmentation problem. MCP fixed it. The 7 million downloads per month and the Linux Foundation stewardship aren't hype — they're evidence that the market recognized something real.
The VentureBeat USB-C comparison is apt. What's also true of USB-C: it transfers data. It does nothing about whether the data is correct.
The Skills problem makes this concrete
Anthropic's announcement of Claude "Skills" — repeatable workflows saved inside Office add-ins and deployed to the whole organization — came with some marketing language that deserves a closer reading.
Anthropic described Skills this way: "Workflows that previously lived in one person's head become one-click actions available to the whole organization."
Right. Now consider what else lives in one person's head: their assumptions. Their outdated pricing. Their interpretation of a policy document from two years ago. Their workaround for a process that officially changed in Q3 but nobody updated the docs.
When those workflows get encoded as org-wide Skills, they stop being one person's problem. They become everyone's one-click error. If a financial analyst saves a Skills workflow built on last quarter's price list, every analyst in the company runs that mistake at the same speed and with the same confidence.
MCP provides the pipe. Skills provides the scale. Neither one checks what's flowing through.
The data is not comforting
This isn't a theoretical concern. According to DataHub's 2026 State of Context Management report, 61% of enterprises delayed AI deployment because they don't trust their data. A separate Liquibase survey found that 64.3% of engineering leaders cite data quality as their top risk in getting AI to production.
These numbers predate MCP's current scale. They reflect what happens when AI agents get limited, careful access to enterprise data. As MCP becomes the standard way for AI to reach every system in the stack, the surface area for bad data to influence AI outputs grows with it.
MCP doesn't create new documents. It gives AI systems machine-scale access to every document that already exists — including the ones that are outdated, contradicting each other, deprecated, or simply wrong. The AI agent market was already projected to hit $52.62 billion by 2030 (MarketsandMarkets). MCP is the infrastructure layer that makes that market function. A bigger market with better connectivity and the same broken knowledge base is a faster path to the same failures.
For a longer look at how document chaos compounds inside agentic systems, we wrote about the 40% agentic AI failure rate last week — the pattern is the same one showing up here at a different layer.
What the coverage missed
Every article about MCP's arrival as universal standard treated connectivity as the finish line. The Verge focused on productivity. VentureBeat covered infrastructure adoption. No outlet asked what happens after the pipe works.
The Manufact co-CEO framed it clearly: "Software products are already being accessed by — and will be accessed mainly by — AI agents, or by users through chat interfaces." His point is about the inevitability of MCP adoption. It's correct. There's a corollary he didn't mention: if software is accessed mainly by AI agents via MCP, then the quality of what those agents retrieve determines everything about what they do.
The connectivity problem was always solvable with engineering. Clean connector interfaces, standardized protocols, a well-funded open-source foundation. MCP solved it. The knowledge accuracy problem is different. It's not an engineering problem. It's an ongoing operational problem: documents decay, contradict each other, go out of date, and get updated in one place but not others. No protocol addresses that. It requires active management.
Where Mojar fits
Mojar AI's Knowledge Base Management Agent is built on MCP — which means MCP's success is directly relevant to what we build. As MCP becomes the standard interface between AI agents and enterprise systems, the knowledge layer behind every MCP server becomes the variable that matters. MCP answers the question: can my AI reach my documents? Mojar answers the harder one: are those documents worth reaching? Contradiction detection, real-time knowledge updates, feedback-driven remediation — these aren't features layered on top of an MCP deployment. They're the layer the MCP stack needs to be complete.
This isn't a criticism of MCP. It's the next step. Connectivity is solved. The industry that built it should be proud of what it shipped. The question for 2026 is what comes after.
What to watch
The organizations that move fastest on MCP won't be the ones with the most servers or the most integrations. They'll be the ones whose connected knowledge is worth reading. That's a different kind of infrastructure problem — one that requires different tooling and a different budget line.
Nobody's celebrating the knowledge layer win yet. That's because it hasn't happened yet. But MCP just made it the most important thing left to solve.
Frequently Asked Questions
MCP gives AI agents a single standardized interface to connect to any software, database, or document system. Before MCP, every integration required a custom connector. MCP eliminated that fragmentation — one protocol, every system. That's a genuine infrastructure win. It doesn't verify whether the content those systems contain is accurate, current, or internally consistent.
Connectivity determines whether AI can reach your documents. Accuracy determines whether those documents are worth reaching. Most enterprise knowledge bases contain outdated files, contradicting policies, and superseded procedures nobody ever cleaned up. MCP gives AI agents machine-scale access to all of it. The garbage-in problem scales with the pipe.
The knowledge layer sits between connectivity (MCP, APIs, integrations) and the AI model. It's responsible for keeping documents accurate, current, and internally consistent — detecting contradictions, updating outdated content, and remediating errors before they flow into agent responses. It's the layer the MCP stack is missing.